Self-Supervised & Contrastive Learning
How models learn strong features from unlabeled data: contrastive losses, SimCLR, BYOL, masked autoencoders, DINO through DINOv3, and JEPA's predict-the-representation idea.
Before this: Deep Learning and Computer Vision, including vision transformers. Multimodal & Vision-Language Models covers CLIP, the image-text side of contrastive learning.
-
10 min
Step 1What self-supervised learning is
Learn features from unlabeled data with training targets made from the data itself.
What Is Self-Supervised Learning and Why Care?Boris Meinardus
-
1h 15m
Step 2Self-supervised learning in CS231N
Stanford lecture tour from pretext tasks to contrastive and masked methods.
Stanford CS231N | Spring 2025 | Lecture 12: Self-Supervised LearningStanford Online
-
5 min
Step 3Contrastive learning in five minutes
The core idea: similar pairs close together, dissimilar pairs far apart.
Contrastive Learning - 5 Minutes with CyrillCyrill Stachniss
-
20 min
Step 4Contrastive loss
The loss behind that idea, worked through step by step.
Contrastive Loss : Data Science Basicsritvikmath
-
15 min
Step 5SimCLR
Two augmented views of one image should match; other batch images are negatives.
Contrastive Learning with SimCLR | Deep Learning AnimatedDeepia
sthalles/SimCLR★ 2.5k -
34 min
Step 6BYOL: no negative pairs
An online network predicts a slow momentum target's embedding, without negatives.
BYOL: Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning (Paper Explained)Yannic Kilcher
lucidrains/byol-pytorch★ 1.9k -
13 min
Step 7Masked autoencoders
Mask most image patches and train a ViT to reconstruct the missing pixels.
Masked Autoencoders Are Scalable Vision Learners – Paper explained and animated!AI Coffee Break with Letitia
-
39 min
Step 8DINO: self-distillation
A student matches a momentum teacher across image crops, with no labels.
DINO: Emerging Properties in Self-Supervised Vision Transformers (Facebook AI Research Explained)Yannic Kilcher
-
8 min
Step 9DINOv2
DINO-style training scaled on curated data, giving general-purpose frozen features.
DINOv2 from Meta AI - Finally a Foundational Model in Computer Vision?AI Papers Academy
facebookresearch/dinov2★ 13.6k -
16 min
Step 10DINOv3
Larger-scale training that keeps dense patch features useful for many tasks.
How AI Taught Itself to See [DINOv3]Jia-Bin Huang
facebookresearch/dinov3★ 11.8k -
8 min
Step 11I-JEPA
Predict the representations of masked image regions instead of their pixels.
I-JEPA from Meta AI - A Human-Like Computer Vision Model | Paper SummaryAI Papers Academy
-
50 min
Step 12V-JEPA
The same feature-prediction idea applied to unlabeled video clips.
V-JEPA: Revisiting Feature Prediction for Learning Visual Representations from Video (Explained)Yannic Kilcher
facebookresearch/jepa★ 4.2k
Build it
Watching isn't learning. Pick one and put it on GitHub.
- Pretrain SimCLR on CIFAR-10 without labels, then train a linear classifier on the frozen features and compare it with a supervised baseline.
- Use frozen DINOv2 features for k-nearest-neighbor classification on a small image set you collect, and plot the embeddings in 2D.
- Show PCA of DINOv2 patch features as colors on your own photos, and point out where objects separate without any labels.